Glenn Gow, The Scaling Executive Coach and host of The Scaling Executive Podcast, sat down with James Hotson — co-founder of Katara AI and a serial company builder whose exits include two acquisitions by Constant Contact and Gladly — to draw a clear operational map for CEOs navigating the agent economy. Hotson’s core claim: most companies are still at level one of agentic automation, automating isolated tasks with logic and rules, when the real competitive advantage comes from redesigning entire workflows around what AI agents can do natively. CEOs who understand that distinction — and start building the infrastructure to act on it — will outpace those still asking whether AI agents are relevant.
This episode is for CEOs of scaling companies who want to move past AI experimentation and understand how to deploy agents operationally, from workflow redesign to internal knowledge infrastructure to what managing an AI-augmented team actually requires.
Key Takeaways
- Most companies are at level one of AI agent deployment — automating isolated repetitive tasks with logic and rules — while the real opportunity is level two: redesigning entire workflows from scratch around what agents and humans each do best.
- The difference between AI agents and no-code tools like Zapier is reasoning. Zapier executes if/then logic. An agent reasons toward a goal, handles complex inputs, and can route, partially answer, or escalate without a human scripting every branch.
- CEOs who deploy an internal AI knowledge base — a private, permission-structured corpus that employees can query in real time — eliminate one of the most common bottlenecks in scaling companies: employees waiting for answers that are already written down somewhere.
- Persistent memory in AI systems, like long and short memory in the human brain, must be structured in layers: a global company knowledge base, function-specific knowledge bases, and externally facing corpora — each with its own access controls.
- Hotson projects that managing AI agents will consume 20 to 30 percent of a CEO’s operational attention within the near term. The skill set required — agent metrics, model cost optimization, data privacy decisions — does not exist in standard CEO training and must be built deliberately.
Why Most CEOs Are Running Level One AI When They Should Build Toward Level Two
James Hotson, co-founder of Katara AI, draws a line between two stages of AI agent adoption that most CEOs conflate. Level one is what most companies are already attempting: identify tedious, time-consuming tasks, drop in an agent with a reasoning model, and reduce the hours a human spends on that task. A customer support bot that classifies inbound tickets as level one or level two and routes them to the right person is a level one deployment. It is useful. It cuts workload. But it does not change the structure of how work gets done.
Level two is what Hotson believes too few companies are doing: stepping back from the existing process entirely and asking what an AI-native version of that process would look like, then designing the workflow from that starting point and determining where humans fit in. The distinction matters because level one automation optimizes a process built for humans. Level two creates a process built for the collaboration between agents and humans — which often looks structurally different.
Hotson points to direct-to-consumer e-commerce as a clear example. In a traditional DTC operation, the fulfillment process starts with a human and passes through multiple humans before a package reaches a customer. A level one approach automates individual steps — routing labels, flagging exceptions, triggering emails. A level two approach rethinks the entire fulfillment chain: what if an agent handled routing from order placement through warehouse dispatch, with a robot executing the physical pick, pack, and ship, and humans only intervening at decision points the agent flags? The result is not an optimized process. It is a structurally leaner operation where human labor concentrates entirely on judgment calls rather than execution steps — reducing the total headcount required for fulfillment while increasing the decisions-per-person ratio.
CEOs who recognize this distinction have a structural question to ask their leadership teams: for our most expensive operational workflows, are we automating steps inside a human process, or are we redesigning the process for an agentic world? Those are not the same initiative, and they do not produce the same results. As Hotson puts it: “I don’t see enough companies doing that really well right now. I see a lot of companies doing that level one.”
What Separates AI Agents from Workflow Tools Like Zapier — and Why the Difference Matters for Scaling
CEOs who have already built no-code automation inside tools like Zapier, Make, or N8N sometimes assume they understand what agents do. Hotson draws a sharp line. Zapier and its equivalents are logic and rules engines. You script the branches. You tell the tool: if this happens, do this. Every path is predetermined. The tool executes. If a situation arises outside the scripted branches, the tool breaks or routes to a human by default.
AI agents operate on reasoning. You give the agent a high-level goal. The agent determines how to reach it. Hotson uses a math analogy that makes this concrete: ask what two plus two equals, and any system gives you four in two seconds — that is pure memory and rules. Ask what 37 times 17 equals, and you need computation. The answer requires reasoning behind the output, not just retrieval of a stored rule. That is the difference between a Zapier workflow and an agent. Zapier handles two plus two. An agent handles 37 times 17.
The operational implication is that AI agents belong on tasks where the input is variable, the right response is not fully predictable in advance, and the cost of a wrong response is low enough to tolerate partial answers. A customer support bot that identifies whether an inbound message is a level one or level two support ticket and routes it — or partially answers it — is doing work that a Zapier workflow cannot do reliably, because the agent must read the content of the message, evaluate its complexity, and make a judgment call.
Hotson flags an important nuance: reasoning is expensive. LLM inference costs money, and not every task requires it. CEOs building agentic infrastructure should map their workflows and separate the two-plus-two tasks from the 37-times-17 tasks. Automate the former with logic. Deploy agents on the latter. Running a high-grade reasoning model on every task in a workflow is inefficient and unnecessary — and optimizing that cost allocation will become a core operational skill for CEOs in the near term.
How CEOs Build an AI Knowledge Base That Actually Reduces Team Bottlenecks
One of the highest-leverage internal applications Hotson identifies is the AI knowledge base: a private, secure corpus of company data that employees can query in real time, structured with layered access permissions so the right people reach the right information.
The problem it solves is one most scaling companies recognize. Employees are blocked. They have a question. The answer exists somewhere in a Notion doc, a Slack thread, a Google Drive folder, or an onboarding deck from 18 months ago. But finding it takes time, and waiting for a human who knows the answer takes longer. A well-built internal knowledge base eliminates both delays. The employee queries the knowledge base, gets an immediate answer, and keeps moving.
At Wire Network, a single customer feedback conversation with Katara AI produced a working product update within one week. A customer named Nick told Hotson that a specific setup task in the Katara product was taking him 20 minutes and suggested a feature that would let him upload a list of questions and answers and have them sync automatically. Katara shipped that functionality within seven days. That same feedback loop — direct conversation producing immediate, specific improvement — is what a well-maintained internal knowledge base enables for a team asking operational questions.
The architecture Hotson recommends has three layers. A global knowledge base holds everything under the company: policies, processes, product documentation, history. Function-specific knowledge bases serve sales, marketing, HR, and the executive team separately, with access provisioned accordingly. An external-facing corpus governs what the company wants AI-driven conversations to look like outside the organization. Each layer requires deliberate curation — the volume and specificity of what goes in determines the quality of what comes out. Too much data, and the model gets confused. Too little, and answers stay generic. Hotson describes curating that balance as an art and science Katara is actively working through.
Persistent memory is the component that makes the architecture durable. Hotson frames it the same way a human brain works: “You want long memory and short memory, just like a human brain. Maybe you also want to layer in multiple corpuses of data — a global knowledge base for everything under ACME Co., and then individual knowledge bases for sales, marketing, HR, executive team.” A knowledge base without persistent memory is a static document library. A knowledge base with persistent memory learns from queries, retains context across sessions, and becomes more useful over time.
What CEOs Must Learn to Manage AI Agents as a Core Operating Responsibility
Hotson is direct about where he sees the CEO role heading. Managing AI agents and agentic workflows will not be a delegated IT function. It will consume 20 to 30 percent of a CEO’s operational attention, and the decisions involved require judgment that only the CEO can provide.
The decisions are not simple. Which agents should run on the highest-grade models, and which should use smaller, cheaper models optimized for their specific tasks? What are the right metrics to evaluate whether an agent is performing well, given that the industry has not yet settled on standard agent performance benchmarks? When a CEO builds a custom agent inside a platform like OpenAI, what data privacy assumptions are they making, and do those assumptions hold — or should that agent run on local infrastructure to avoid training future models on proprietary company data?
These are technology decisions with strategic consequences. A CEO who cannot make them intelligently — or who defaults all of them to a CTO without understanding the cost and risk structure — is operating blind in a part of the business that will define competitive position over the next five years.
Hotson’s framing is not doom and gloom. He is explicit that AI is not eliminating jobs. It is changing roles. People who previously spent hours on repeatable, rule-bound tasks will shift to work that requires judgment, relationship, and creativity — the work that agents cannot do reliably yet. But that shift requires CEOs to redesign role expectations, measure agent performance with the same rigor they apply to human performance, and make the technology decisions that determine how well the agents and humans collaborate. As Hotson frames it: “Whether that’s 20 percent or 30 percent of your business and your time getting the right agentic workflows, the right AI agents is going to be a major shift in how CEOs operate companies.”
The CEO who treats agentic infrastructure as a technology project owned by an engineering team will be behind. The CEO who treats it as a core operating system — one they understand, govern, and continuously optimize — will compound the advantage every quarter.
The Scaling Executive Framework: How James Hotson Thinks About AI Agents for CEOs
| Principle | What It Means in Practice | Named Evidence from This Interview |
| Separate logic tasks from reasoning tasks before deploying agents | Agents running reasoning models are expensive — only deploy them where variable inputs require judgment, not where rules and memory will do | Katara AI’s customer support routing uses agent reasoning to classify inbound messages — enabling the system to handle classification and partial answers that Zapier-style logic tools cannot produce, because no pre-scripted branch covers every message variation |
| Don’t automate your existing process — redesign it for agents | Level one automation optimizes a human process; level two rebuilds the workflow from the ground up around what agents and humans each do natively | A level two DTC fulfillment redesign concentrates all human labor on judgment-only decision points — eliminating humans from pick-pack-ship execution entirely and reducing the total headcount the fulfillment operation requires |
| Build a layered knowledge base with persistent memory before you need it | A global corpus plus function-specific corpora with layered access controls eliminates team bottlenecks and compounds in value as it learns from queries | At Wire Network, a single Katara AI customer conversation produced a shipped product feature within one week — compressing a feedback-to-delivery cycle that would typically span months into seven days |
| CEOs must own agent governance, not just delegate it | Model selection, cost optimization, data privacy decisions, and performance metrics for agents require CEO-level judgment — these are strategic, not just technical | Hotson projects that CEOs who cede these decisions to engineering teams will lose competitive ground as agentic infrastructure defines operational positioning over the next five years — a consequence he describes as compounding every quarter for those who govern it directly |
Quotes from This Episode
- “You want long memory and short memory, just like a human brain. Maybe you also want to layer in multiple corpuses of data — a global knowledge base for everything under ACME Co., and then individual knowledge bases for sales, marketing, HR, executive team.” — James Hotson, Co-founder, Katara AI
- “I don’t see enough companies doing that really well right now. I see a lot of companies doing that level one.” — James Hotson, Co-founder, Katara AI
- “Whether that’s 20 percent or 30 percent of your business and your time getting the right agentic workflows, the right AI agents is going to be a major shift in how CEOs operate companies.” — James Hotson, Co-founder, Katara AI
- “When a CEO builds a custom agent inside a platform like OpenAI, what data privacy assumptions are they making, and do those assumptions hold — or should that agent run on local infrastructure to avoid training future models on proprietary company data?” — James Hotson, Co-founder, Katara AI
Frequently Asked Questions
What is the difference between AI agents and tools like Zapier or Make?
Zapier, Make, and N8N are logic and rules engines — they execute scripted if/then branches you define in advance, and they break when a situation falls outside those branches. AI agents operate on reasoning: the agent receives a high-level goal and determines how to reach it, handling variable inputs that no pre-scripted path could anticipate. James Hotson, co-founder of Katara AI, frames it this way: rules-based tools handle two plus two, where memory retrieves a known answer. Agents handle 37 times 17, where inference must compute the answer from first principles. The operational consequence is that agents belong on tasks where the right response is not fully predictable — like classifying and routing customer support tickets — while Zapier belongs on deterministic, high-volume tasks where every branch is known.
How should a CEO structure an internal AI knowledge base for a scaling company?
James Hotson, co-founder of Katara AI, recommends a three-layer architecture: a global knowledge base holding all company data, function-specific knowledge bases for departments like sales, marketing, HR, and the executive team — each with its own access controls — and an externally facing corpus governing AI-driven customer interactions. Persistent memory is what makes the system durable: long memory holds foundational knowledge, short memory retains recent context, and the system becomes more useful as it learns from queries. Hotson cautions that curation is the hard part — too much data confuses the model, too little produces generic answers — and treats calibrating that balance as an ongoing discipline, not a one-time setup.
What does managing AI agents actually require from a CEO?
James Hotson projects that managing AI agents and agentic workflows will consume 20 to 30 percent of a CEO’s operational attention in the near term. The decisions involved include which agents should run on high-grade reasoning models versus smaller cost-optimized models, what metrics signal agent performance given that industry benchmarks do not yet exist, and whether to build agents inside third-party platforms or run them locally to protect proprietary training data. These are strategic decisions with compounding consequences — not IT decisions to delegate. CEOs who treat agentic infrastructure as owned by engineering will lose ground to those who govern it directly.
CEOs Work with Glenn Gow to Deploy AI Agents and Build Operational Leverage
Glenn Gow is The Scaling Executive Coach — he coaches ambitious executives into the CEO seat and CEOs into successful exits. With 25 years as a CEO and 5 years in venture capital, Glenn helps leaders scale their companies by scaling themselves first. If this conversation was useful, you can apply for executive coaching with Glenn Gow or apply to be a guest on The Scaling Executive Podcast.
